This week we were tasked with “building an interface for conjuring an environment rather than traveling through it”. This week we were also introduced to the machine learning software RunwayML. I decided to train a model on RunwayML based on images of my apartment, and present them as a fluid scene in a p5.Js animation.
I trained my ML model with approximately 1000 photos of my apartment. I got these photos by taking videos of my apartment at different times of the day and breaking those videos into still images with the command line tool FFmpeg.
Because I was using a one-time/free/trial version of RunwayML’s training model, as you can see in the video above, the ML generated images came out incredibly distorted. Were I to try and replicate this project, I would be inclined to spend some money to allow this model to train for longer, to see if I could generate realistic apartment images.
Though I will not share my apartment images or my RunwayML model, I will share my p5 code.